One‑minute snapshot
*Prices are normalized per single H200 even when a provider sells only 8‑GPU nodes. U‑S region pricing, on‑demand only (no spot, reserved, or contract rates).
Methodology – why you can trust these numbers
<ul><li><strong>On‑demand only.</strong> We excluded capacity reservations longer than 14 days, reserved instances, and spot/pre‑emptible offers.</li><li><strong>Same silicon.</strong> Every row is a 141 GB NVIDIA H200 (SXM or PCIe).</li><li><strong>Public price lists only.</strong> Figures come straight from each provider’s pricing page on 25 July 2025.</li><li><strong>US regions, USD.</strong> Regional variation can add 5‑20 percent; those are ignored for apples‑to‑apples comparison.</li></ul>
A100 vs H200 cost benchmark across generations
Bottom line: two hours on Thunder Compute’s A100 costs less than 15 minutes on Azure’s H200—and still buys roughly 13 × more runtime per dollar than hyperscaler H200s.
Takeaways for developers
<ul><li><strong>H200 premiums remain steep.</strong> Even after AWS’s June price cut, the cheapest hyperscaler H200 hour is >5 × Thunder Compute’s A100.</li><li><strong>Specialist clouds narrow the gap.</strong> Lambda, RunPod, Jarvislabs, and Vast.ai all sit in the $2–4 range, but Thunder’s A100 is still ~2‑5 × cheaper.</li><li><strong>Choose H200 only when you must</strong>—massive models that overflow 80 GB VRAM or long‑context inference. For prototyping, fine‑tuning, and most training, A100 80 GB wins on ROI.</li><li><strong>Thunder Compute roadmap.</strong> We don’t offer H200 nodes yet; today you can launch A100 80 GB at $0.78/hr (one‑click VS Code, per‑second billing, persistent volumes, live hardware swaps).</li></ul>
Bookmark this page—we refresh the numbers quarterly. Meanwhile, spin up an A100 or H100 on Thunder Compute and keep more budget for your model weights.